Consensus Alignment Improves Child-Speech ASR Training Data

Long CHILDES recordings pair valuable natural child-adult interaction audio with transcripts whose utterance timestamps can be noisy, incomplete, or misaligned. That prevents researchers from reliably extracting supervised speech segments: a transcript may be correct while its stated location in the recording is not. The paper introduces BEACON, Boundary Estimation via Alignment CONsensus, as an automatic timestamp-curation pipeline that treats agreement among several ASR alignments as evidence that a clip boundary is trustworthy.

Core Contribution

BEACON does not train a new recognizer or require new manual boundary labels. Instead, it uses multiple existing ASR models to generate word-level timestamps, aligns each model output to a trusted human transcript, and estimates utterance boundaries from their consensus. This is a practical distinction from relying on a single forced aligner: disagreement becomes a signal for withholding uncertain clips rather than silently treating every timestamp as correct.

The resulting releases serve two different purposes. The general-purpose corpus retains 413.3 hours and the original CHAT annotations; the stricter ASR-training subset contains 283 hours after transcript normalization and WER-based mismatch filtering. The authors frame the recipe as corpus-agnostic, provided that a long recording has a trusted transcript but unreliable or absent timestamps.

Technical Approach

For each utterance, BEACON aligns ASR word timestamps to the reference transcript, then combines candidate start and end times using consensus voting. An ablation on RSR child speech and AMI meetings finds that four-model ensembling gives lower mean absolute boundary error than any single model: 0.44 s on RSR and 1.11 s on AMI, compared with individual-model ranges of 0.97–3.49 s and 1.33–3.81 s, respectively. The benefit is strongest where one system's alignment error can be offset by another's.

The pipeline includes two material curation choices. First, adjacent same-speaker child utterances separated by at most 1 s are merged, with 0.5 s padding at each side. This addresses very short items such as “yeah” and “mhm,” whose boundaries are difficult to localize in isolation. Second, the ASR subset removes clips with an insertion or deletion rate above 0.25 when a fixed ASR transcript is compared with the normalized reference. These filters make the training set smaller, but aim to make its audio-text pairs more dependable.

Figure 1 summarizes the BEACON pipeline from multi-model timestamp generation through transcript alignment, voting, filtering, and release construction.

Results and Analysis

On CHILDES, the raw release totals 605.9 hours but has a bounded proxy WER of 63.7%. BatchAlign2 variants retain 571.6–603.8 hours with 60.9–62.3% WER. FASA is more selective, retaining 167.6 hours at 49.7% WER. BEACON occupies a different point on that trade-off: 432.4 hours at 51.3%, or 413.3 hours at 48.5% after turn merging. The merged BEACON release therefore gives up about 19 hours relative to its unmerged form while improving the proxy score by 2.8 points; it retains far more audio than FASA while approaching FASA's proxy quality.

The downstream test is more persuasive than the proxy alone. Fine-tuning Whisper, Parakeet, and Canary on the curated set is evaluated on four held-out child-speech benchmarks: RSR, MyST, OGI Kids, and CMU Kids. For Whisper, the authors report WERs of 21.4, 8.5, 24.2, and 18.3, versus 25.2, 12.8, 69.3, and 16.3 without fine-tuning, for a 19.5% mean relative reduction. Parakeet improves by 3.6% on average and Canary by 5.6%; FASA-based fine-tuning yields 12.4%, 1.3%, and 4.6% reductions for those same model families.

That pattern supports the central claim, with an important qualification. Better timestamp selection translates into useful transfer across several recognizers, particularly for Whisper and the OGI benchmark. But gains are uneven: Whisper's CMU Kids WER rises from 16.3 to 18.3, Parakeet's MyST score rises from 10.5 to 10.7, and Canary's MyST score rises from 9.5 to 9.7. The evidence supports BEACON as a data-curation method, not as a guarantee that every child-speech domain or recognizer benefits uniformly.

Limitations

The authors explicitly note that consensus discards clips when models disagree, including some correctly labeled clips whose audio is difficult for the selected ASR systems. The retained set can therefore favor easier speech. Hyperparameters were chosen through light manual inspection and may not transfer unchanged to other recording conditions or speakers. The reported downstream evidence is confined to English CHILDES-derived data and four child-speech benchmarks, so the broader corpus-agnostic claim has not yet been tested across languages or transcription conventions.